This dissertation addresses the problem of efficient algorithms for recognizing objects based on extracted features, by matching against large databases of characteristic signature patterns.;We review the field of object recognition in image understanding and present a theory of feature-based matching. Our formulation is in the framework of Bayesian decision theory. The objective is to provide a sound scoring mechanism to measure the similarity of the match between a collection of image features and a collection of model features incorporating uncertainty and statistical model variability. The scoring should be robust in the presence of noise, obscuration and complex model and signature variation.;Our applications are in the domain of recognition theory for image understanding, sensors exploitation, and next-generation automatic target recognition. We compare match scoring mechanisms in use in the object recognition field with several new scoring measures.;We discuss algorithmic implementation issues, and we present extensive experimental results focusing on complexity, scalability and robustness of the techniques. |